Defining Point-Set Surfaces - Computer Science @ UC Davis

advertisement
Defining Point-Set Surfaces
Nina Amenta∗
University of California at Davis
Abstract
The MLS surface [Levin 2003], used for modeling and rendering
with point clouds, was originally defined algorithmically as the output of a particular meshless construction. We give a new explicit
definition in terms of the critical points of an energy function on
lines determined by a vector field. This definition reveals connections to research in computer vision and computational topology.
Variants of the MLS surface can be created by varying the vector
field and the energy function. As an example, we define a similar surface determined by a cloud of surfels (points equipped with
normals), rather than points.
We also observe that some procedures described in the literature
to take points in space onto the MLS surface fail to do so, and we
describe a simple iterative procedure which does.
1
Yong Joo Kil†
University of California at Davis
the relationship of extremal surfaces and implicit surfaces. As we
discuss in Section 5, there is an implicit surface containing every
extremal surface, including the MLS surface. This can be quite
useful, particularly for defining normals precisely.
Introduction
Because of improved technologies for capturing points from the
surfaces of real objects and because improvements in graphics hardware now allow us to handle large numbers of primitives, modeling
surfaces with clouds of points is becoming feasible. This is interesting, since constructing meshes and maintaining them through
deformations requires a lot of computation. It is useful to be able
to define a two-dimensional surface implied by a point cloud. Such
point-set surfaces are used for interpolation, shading, meshing and
so on.
David Levin’s MLS surface [Levin 2003] has proved to be a very
useful example of a point-set surface. Levin defined the MLS surface as the stationary points of a map f , so that x belongs to the
MLS surface exactly when f (x) = x. This definition is useful but it
does not give much insight into the properties of the surface.
In Section 2 we give a more explicit definition of the MLS surface, based on an energy function e and an vector field n. Changing
e and n produce other similar surfaces; we call them all, including
the MLS surface, extremal surfaces.
As we discuss in Section 4, extremal surfaces are not new. The
name is adopted from Medioni and co-authors [2000], who used
them for surface extraction from very noisy point clouds, among
other applications in computer vision. Their definition has to be
extended slightly in order to include the MLS surface, but the idea
is essentially the same. Edelsbrunner and Harer [to appear] have
recently described Jacobi surfaces, which are also closely related
and which have stronger mathematical properties.
Adamson and Alexa [2003a] describe an implicit surface which
is a useful “relative” of the MLS surface. This raises the question of
∗ Supported by NSF ACI-0325934, CCR-0098169, and CCR-0093378.
e-mail: amenta@cs.ucdavis.edu
† Supported by NSF CCR-0093378. e-mail:kil@cs.ucdavis.edu
Figure 1: An example of an extremal point-set surface which takes surfels, rather
than points as input. The sparse and non-uniformly distributed set of weighted surfels
on the left implies the surface on the right.
As an example of another extremal surface construction, we
describe in Section 7 a point-set surface for surfels, input points
equipped with normals. Normals are available when converting a
model from a mesh or implicit surface to a point cloud, and they
are generally computed as part of the process of cleaning up point
clouds produced by laser range scanners or other 3D capture technologies. Figure 1 shows an example of our point-set surface for
surfels.
MLS surfaces have been used widely in the last few years. The
seminal graphics paper [Alexa et al. 2003] using the MLS surface
for point-set modeling and rendering was followed by work on progressive MLS surfaces [Fleishman et al. 2003], ray-tracing [Adamson and Alexa 2003b], and surface reconstruction [Xie et al. 2003;
Mederos et al. 2003]. The MLS surface is used in several features
of PointShop 3D, an excellent open-source point-cloud manipulation tool [Zwicker et al. 2002; Pauly et al. 2003]. We have used
PointShop as our implementation platform.
These papers describe two slightly different procedures for taking a point x in the neighborhood of the point cloud onto the MLS
surface, one which first appeared in an earlier, widely circulated
manuscript version of Levin’s paper, and a more efficient “linear”
version used in PointShop. Neither procedure actually produces a
point of the MLS surface. We give a simple procedure which does,
together with a short proof, in Section 6, and we discuss the theoretical problems with the earlier procedures in Appendix A. The
final version of Levin’s paper on the MLS surface [Levin 2003]
(currently available on his Web site) contains a new projection procedure, different from ours, which also produces points on the sur-
face.
2
Surface definition
We begin with the now-standard definition of the MLS surface,
given in the early manuscript of Levin’s paper and used in a variety of contexts as mentioned above. The MLS surface for a point
cloud P ⊆ IR3 is defined as the set of stationary points of a certain
function f : IR3 → IR3 . An optional polynomial fitting step, which
we omit here, can be applied after the map f .
The procedure for computing f (r), described above, minimizes
eMLS (~a,t) over the three-dimensional domain S2 × IR. Therefore
using the new notation we do not minimize over all of the fivedimensional domain IR3 × IP2 , but over a three-dimensional subset:
the set Jr of values (x, a) such that x = r + t~a for some t ∈ IR and
orientation ~a of a. This means that in Jr , every point x ∈ IR3 other
than r is paired with the direction a of the line through x and r; the
singular point r is paired with all directions a. Different values of r
produce different values of f (r) because each choice of r implies a
different domain Jr over which eMLS is minimized.
3
a
Explicit definition and generalization
Now we want to give an explicit version of the MLS surface definition. We begin by defining an (unoriented) vector field:
r
t
n(x) = argmina eMLS (x, a)
x = r + ta
H
Figure 2: The MLS energy function eMLS (~a,t) sums up the weighted distances from
the fixed input points in P to the plane with normal ~a through the point x = r +~at. The
weight on an input point pi ∈ P, denoted here by its shade of grey, is a function of the
distance from pi to x.
Given an input point cloud P and a point r in a neighborhood of
P, the energy of the plane with normal ~a passing through the point
x = r + t~a, where t is the distance from r to the plane, is defined to
be:
eMLS (~a,t) =
∑ (h~a, pi i − h~a, r + t~ai)
pi ∈P
2
θ (p + t~a, pi )
where the weighting function θ is any monotonic function, usually
a Gaussian:
2
θ (x, pi ) = e
This is the normal to the plane through x ∈ IR3 which is a local
best-fit to the point cloud P. Since fixing x fixes the weights, the
energy function is a quadratic function of a and the minimal direction is usually unique. It can be found efficiently as the smallest
eigenvalue of a three-by-three matrix [Alexa et al. 2003]. Where
there are two or three smallest eigenvalues, n is not well-defined.
The sets of points with multiple smallest eigenvalues form surfaces
which divide space into regions, within each of which n is a smooth
function of x.
S
n(x)
x
x
−d (x−pi )
h2
Here h is a constant scale factor, and d() is the usual Euclidean
distance between points. See Figure 2. The energy measures the
quality of the fit of the plane to P, where pi ∈ P is weighted by its
distance from x = r + t~a.
The local minima of this energy function, over S2 × IR (S2 is the
space of directions, the ordinary two-sphere), occur at a discrete set
of inputs (~a,t), each corresponding to a point x = r + t~a. Of these,
f (r) is defined to be the x nearest to r. The stationary points of this
map f form the MLS surface.
We can get some additional insight into this energy function by
restating it using different notation. First, we write it as a function
of ~a and the point x = r + t~a, rather than as a function of ~a and
t. Second, we notice that the plane and the weights determined
by the parameters (x,~a) are the same as those determined by the
parameters (x, −~a), so we can write eMLS as a function of x and an
unoriented direction vector a. This gives us:
eMLS (x, a) =
(1)
∑ (ha, pi i − ha, xi)2 θ (x, pi )
pi ∈P
(although the inner product is not defined for unoriented direction
vectors, we use this notation since we can evaluate the function
using either ~a or −~a and get the same result). In this new form, the
domain of the function is now IR3 × IP2 , where IP2 (the projective
two-sphere) is the space of unoriented directions. The new notation
makes it clear that eMLS (x, a) is independent of r, which will help
us find a non-algorithmic characterization of the points of the MLS
surface in the following section.
Figure 3: To see if a point x belongs to the MLS surface, we consider eMLS on the
line `x,n(x) . Keeping n(x) fixed, we vary y along the line. If y = x is a local minimum
of eMLS (y, n(x)), then x belongs to the MLS surface. Using different functions for n(x)
and e(x, a) gives variants of the construction, which we call extremal surfaces.
Now we give an explicit characterization of the MLS surface;
in effect, we describe how to recognize whether a point x ∈ IR3
belongs to the MLS surface. This characterization is illustrated in
Figure 3. Let `x,n(x) be the line through x with direction n(x). We
adopt the notation arglocalminy to refer to the set of inputs y producing local minima of a function of variable y.
Claim 1 The MLS surface consists of the points x for which n(x) is
well-defined, and for which
x ∈ arglocalminy∈`x,n(x) eMLS (y, n(x))
Proof : First we argue that every x on the MLS surface has this
property. Such a point corresponds to a pair (x, a) which is a local minimum in its own set Jx . The set A = {(x, a)}, where x is
fixed and a ranges over all possible directions, is a subset of Jx , so
(x, n(x)) is a local minimum in A. Since n(x) is well defined, n(x) is
the unique global minimum in A and we have a = n(x). The set of
pairs L = {(y, n(x)) | y ∈ `x,n(x) } is also a subset of Jx , so (x, n(x))
is also a local minimum in L.
Now we want to show that any x which has the property in the
Claim belongs to the MLS surface. We need therefore to show that
(x, n(x)) is a local minimum of eMLS in the set Jx .
Consider any direction m 6= n(x). Since n(x) is defined as the
direction producing the unique minimum over all pairs (x, a), we
have eMLS (x, m) > eMLS (x, n(x)). The function eMLS is continuous,
so there is some distance ε (m) such that for all y ∈ `x,m with
d(x, y) < ε (m), we have eMLS (y, m) > eMLS (x, n(x)). Also there is
some ε (n(x)) such that for all y ∈ `x,n(x) with d(x, y) < ε (n(x)),
eMLS (y, n(x)) > eMLS (x, n(x)). Let ε be the minimum of ε (a) over
all directions in a ∈ IP2 . Then (x, n(x)) is a local minimum in the
subset of Jx consisting of pairs (y, a) with d(x, y) < ε . 2
We can generalize the MLS construction by considering alternatives for the two functions n and eMLS . We can use any function
n(x) to assign directions to points in space, and any function e(x, a)
to specify the energy of elements of IR3 × IP2 . There is no reason
why the definition of n has to be related to the definition of e, as it
is for the MLS surface. We define an extremal surface as follows.
manifold. This does not extend to higher dimensions; for instance a
Jacobi 3-surface in IR4 can be generically non-manifold, indeed at
the singular points at which n would be undefined. They prove that
Jacobi k-surfaces in IRd , for d > 2k − 2, are manifolds, and they
give a robust algorithm for extracting Jacobi surfaces from functions given on a tetrahedral mesh.
5
Implicit and extremal surfaces
Adamson and Alexa [2003a] defined an implicit surface which they
used for ray-tracing instead of the MLS surface. Their surface has
the form
g(x) =~n(x) · (a(x) − x) = 0
where ~n(x) is an oriented vector field and a(x) is the center of mass
of the input point cloud P as weighted by x.
Definition 1 For any functions n : IR3 → IP2 and e : IR3 × IP2 → IR,
let
S = {x | x ∈ arglocalminy∈`x,n(x) e(y, n(x))}
be the extremal surface of n and e.
4
Extremal surface literature
Not surprisingly, the idea of extremal surfaces is not new. Guy
and Medioni [1997] and Medioni, Lee and Tang [2000] defined extremal surfaces, using functions n : IR3 → IP2 and s : IR3 → IR, to
define the set {x | x ∈ arglocalminy∈`x,n(x) s(y)}. Our definition is a
a little more general than theirs in that their function s(x) = e(x, a)
does not require the parameter a. In a series of papers, they used
extremal surfaces for (among other things) surface reconstruction
from very noisy point clouds. Their functions n and s are different
from the MLS energy function, and require completely different
computational techniques. They represent n and s simultaneously
with a tensor, and use tensor operations to smooth them. This tensor
voting is performed on a voxel grid. The extremal surface is then
extracted from the grid with the marching cubes algorithm. Most
of their work focuses on the difficult problem of designing of good
tensor functions.
Edelsbrunner and Harer [to appear] define Jacobi surfaces in IR d .
To keep this discussion simple, we give their definition for the special case of two-surfaces in IR3 . The input is three Morse functions
f1 (x), f2 (x), f3 (x) (intuitively, a Morse function is one whose isosurfaces are generic; it is everywhere twice-differentiable, its Hessian matrix is everywhere non-singular, and no two critical points
have the same function value). Jacobi surfaces are symmetric with
respect to the order of the input functions, so that for instance if we
exchange f 1 and f3 , we get the same Jacobi surface.
The intersections of the level sets of f 1 and f2 divide IR3 into a
family of curves. The Jacobi surface S is defined as the set of critical
points of f3 over each of these curves. Every point x ∈ IR3 belongs
to one such curve, and we let n(x) be the tangent direction. Every
critical point q of f 3 on the curve containing x is a critical point of f 3
on the tangent line `x,n(x) as well, so this is similar to an extremal
surface with f 3 (x) = s(x). The main difference is that all critical
points, rather than just minima, are taken. Another difference is that
points at which n is undefined (because the intersection of the level
sets consists of a single point instead of a curve) are included in the
Jacobi surface (this is related to the symmetry of the definition).
With these points included as part of the surface, it seems
likely that these singularities in the vector field n might cause nonmanifold singularities in the surface S, for instance points at which
multiple sheets of surface come together. Edelsbrunner and Harer
show, however, that a Jacobi two-surface in IR3 is generically a
Figure 4: Streamlines (red) of a vector field n(x), and iso-contours (blue) of an
energy function s(x). The heavy blue line is the extremal surface determined by n and
s, running neatly along the “valley” in the energy landscape and passing through the
minima of s. The streamlines of n and the iso-contours of s are tangent at the surface
points. Here n and s were computed using the point-set surface for surfels introduced
in Section 7; the input surfels are shown as black diamonds, with the long diagonal
pointed in the direction of the surfel normal.
Figure 5:
The red streamlines indicate a constant vector field ~n. The blue isocontours show an energy function s again determined by the set of input surfels (black
diamonds), here meeting at a sharp corner. There are two valleys in the energy landscape meeting to form a third valley. The implicit surface g(x) = ~n(x) · ∇s(x) = 0
includes both minima in the extremal surface definition (heavy blue curve) and also
maxima (green curve). The extremal surface (heavy blue curves) appears to have a
junction but is actually composed of two manifold components. Using the best-fitting
plane to determine the vector field ~n, as defined for the MLS surface in Equation 1,
produces a similar structure near the sharp corner, but the somewhat larger picture is
complicated by singularities in the vector field.
They prove that g(x) is a smooth function on any domain on
which ~n is well-defined everywhere, and therefore that the surface
juncture one of the sheets ends at a boundary where the critical
points on `x,n(x) switch from minima to maxima.
Describing an extremal surface as a subset of an iso-surface gives
an analytic expression for its normal. In the case of the MLS surface
this formula includes the derivatives of the weights and is rather
complicated. Note that in general the surface normal at a point x is
different from the direction n(x); see Figure 6.
6
Figure 6:
Surfels distributed on a part of the extremal surface of Figure 8 where the
optimal direction n(x) at a surface point x is very different from the surface normal at
x. On the left, the surfels are oriented by n(x), and on the right by the surface normal.
g(x) = 0 is a manifold on the domain, assuming it avoids critical
points of g. They argue that it does, generically (meaning that if
g(x) = 0 does contain a critical point of g, a small perturbation of
the input fixes the problem).
Just considering a domain within which the vector field~n is welldefined neatly avoids the singularities found in higher dimensions
with Jacobi surfaces. Medioni et al. pointed out [2000] that in
a similar context there is an implicit surface associated with any
extremal surface. We consider a domain U ⊆ IR3 within which not
only is n always defined, but it also allows a consistent orientation,
meaning that adopting this orientation produces a smooth oriented
vector field~n. Since a point of the extremal surface is a critical point
of s on the line `x,n(x) , the directional derivative of s in direction n(x)
has to be zero at x, and we can define the implicit surface
Finding surface points
At least two different procedures for taking arbitrary points in IR3
to the MLS surface have been proposed. The projection procedure
given by Levin [Levin 2003], based on the definition above, was
used in the seminal paper [Alexa et al. 2003]. A somewhat different
procedure, designed so as to avoid numerical optimization, is used
in PointShop [Zwicker et al. 2002]. In this section we give another
simple procedure for generating surface points.
We also give a simple proof of correctness. Neither of the earlier
procedures were rigorously proved correct, and interestingly upon
very close examination neither of them actually produces a point
on the MLS surface, although in practice the error is negligible.
We discuss the technical difficulties with these procedures in Appendix A.
)
1
n(x
x1
S
x2
xn
g(x) =~n(x) · ∇s(x) = 0
That is, at a point x of the extremal surface, n(x) is perpendicular
to the gradient of s, and tangent to the iso-surface s(y) = s(x), for
y ∈ IR3 . This is illustrated in Figure 4. Although it is tempting to
assume that the orientation of n is unnecessary since the zero-set
of g2 (x) would be the same with either orientation, the zero-set of
g2 (x) consists entirely of critical points and so may not be a manifold, particularly where n does not admit a consistent orientation.
Using our more general formulation of extremal surfaces, which
includes the MLS surface, we can define the surface
g(x) =~n(x) · ∇y e(y,~n(x))|x ) = 0
where ∇y e(y,~n(x))|x is the gradient of e as a function of y, keeping
n(x) fixed, and then evaluated at x.
Any of these iso-surface are manifolds over the domain U, so
long as they avoid critical points of g, which they do generically.
Notice that a point x on one of the iso-surfaces might be a maximum
on `x,n(x) rather than a minimum. Taking only the minima, as in
the MLS surface definition, might further eliminate parts of each
surface, as in Figure 5, where the part which is eliminated indeed
should not be included in the point-set surface.
We summarize as follows.
Observation 2 The MLS surface, within a domain where n is welldefined and admits a consistent orientation, is, generically, a subset
of a manifold.
This seems to contradict our experience, for instance when the input
points come from surfaces with sharp corners, in the vicinity of
which two sheets of the MLS surface seem to collapse into one.
Looking Figure 5, however, we see that that just before the apparent
Figure 7: Diagram of the process for taking a point to an extremal surface. Point x 1
moves to a local minimum on the line `x1 ,n(x1 ) , represented by the point at which the
dashed lines meet. This becomes x2 . When the process converges, xn lies on S.
Our process for taking a point onto the extremal surface S implied by n and e is illustrated in Figure 7. At each iteration, we find
n(xi ) and consider the line `xi ,n(xi ) . We search for a nearby local
minimum of e(y, n(xi )) over the set y ∈ `xi ,n(xi ) . The nearest local
minimum becomes xi+1 . Notice that as long as resetting n(xi+1 )
at each new point does not increase e, the energy decreases at every step so that this process is likely to converge. The energy does
indeed decrease for the MLS function and also for any function
e(x, a) = s(x) which does not depend on the direction parameter.
Claim 3 If the procedure above converges, it produces a point of
S.
Proof: At convergence, repeating the procedure for xn produces
the same point xn . Since xn is a local minimum of e(y, n(xn )) within
y ∈ `xn ,n(xn ) , this means that xn ∈ S according to Definition 1. 2
7
Point-set surface for surfels
Now we define a point-set surface which takes as input a set of
surfels rather than a point cloud. Normals are often available, and
using surfels rather than points makes the surface more robust in
the face of both undersampling and of irregularities in the distribution of samples; see Figure 8. Our input surfels are represented
as point-direction pairs (pi , ai ). Following the intuition that n(x)
Figure 8: From left to right, a sparse set of surfels defining a chess-piece. Next, we take a 3D grid of points onto the surface using our point-set surface (with c = 0). In the
center, we find that the points do indeed go to a two-dimensional surface. The MLS surface, as implemented in PointShop, has trouble on this example. Without using the normal
information, the very sparse, non-uniform distribution of points makes the MLS energy function give very good scores to the planes through the vertical rows of points; we show the
grid points as projected by MLS. Finally, at the far right, the complete surface produced by our point-set surface.
should follow the surface normal at the nearest surface point to x,
we compute n as a weighted combination of the nearby surfel normals. If we have oriented normals, we can take a vector average.
n(x) = ∑ ai θN (x, pi )
i
where
θN (x, pi ) =
e−d
2
(x,pi )/h2
∑ j e−d
2 (x,p
j )/h
2
is a normalized Gaussian weighting function.
If the surfels have unoriented normals, we instead use the principal component of the normal vectors, again weighted by θ ; this
is the direction of the line through x which is the weighted leastsquares best fit to the vectors, and it can be computed as the eigenvector of largest eigenvalue of the 3 × 3 covariance matrix B where
b jk = ∑ θN (x, pi ) ai, j ai,k ∀ j, k = {1, 2, 3}
i
and the (x, y, z) coordinates of vector ai are (ai,1 , ai,2 , ai,3 ). This is
not quite as efficient as the vector average, but in either case computing n(x) is faster than minimizing e.
Our intuition is that e is an estimate of unsigned squared distance
function, based on surfel position and normal. We define the distance of x from a surfel as a Mahalanobis distance (like Euclidean
distance but with elliptical rather than spherical unit ball), where
distance in direction ai is emphasized over directions perpendicular
to ai .
dM (pi , ai , x) = h(x − pi ), ai i2 + c || (x − pi ) − h(x − pi ), ai i ai ||2
With the scaling factor c = 1 we have the Euclidean distance from
x to pi , and with c = 0 we have the squared distance from x to the
plane through pi normal to ai . Figure 9 shows the effect of different
choices of the parameter c. Finally we define
e(x, a) = e(x) = ∑ dM (pi , ai , x)θN (x, pi )
Figure 9: The point-set surface produced by six surfels. The constant c in the energy
function is one at the upper center, then halved for each successive image, and finally
zero at the lower right.
minimum of e on the line `x,n(x) , we define q = x + τ n(x), and minimize e(q) as a function of τ . We used an implementation of Brent’s
method for this one-dimensional non-linear optimization from Numerical Recipes in C [1992], similar to Alexa et.al. [2003]. For
efficiency, we used PointShop’s kd-tree to find nearby surfels, and
used contributions only from the nearest 30.
Using θN instead of θ is important since with the simple Gaussian the energy would be effectively zero far from the surface.
When the Gaussian weight on every point is numerically zero we
cannot compute θN . In that case we let θN be one for the nearest
surfel and zero for all others, which is nearly correct.
Our implementation allows input surfels to have variable
weights, so that we can vary their distribution on the surface, as
in Figure 1. This is implemented by storing a separate weight hi
with each surfel.
i
We implemented the procedure for taking points in space to this
extremal surface as part of a plug-in for PointShop. To find the
θN (x, pi ) =
e−d
∑j e
2
(x,pi )/h2i
−d 2 (x,p j )/h2j
definition, but there may be simpler or more robust approaches.
There are many issues to be resolved with respect to procedures
for taking points in space to an extremal surface. Despite the theoretical problems with both Levin’s original procedure and with
the heuristic used in PointShop, both methods have advantages that
ours does not: the projection idea is very elegant, while the heuristic is efficient. The new projection procedure in the final version
of Levin’s paper [Levin 2003] could be applied with any extremal
surface, and may be useful in practice.
While the iso-surface definition of an extremal surface gives an
expression for the normal, it is often complicated, since it includes
the derivatives of the weight functions. Are there extremal surfaces
for which the normals have a particularly simple form? Or is there
some way to use a simpler implicit function, such as that proposed
by Adamson and Alexa, without including parts such as the green
curve in Figure 5?
Finally, it would be good to prove connections between some
extremal surface and the real surfaces from which point samples
are taken. Under what sampling conditions can we guarantee that
the extremal surface defined by a sample point cloud is everywhere
close to the original surface? Under what conditions can we guarantee that there is an isotropy between them?
9
Figure 10: Using our surfel point-set surface definition to fill in a set of samples for
rendering. The input data is the vertex set of a polygonal model. We produced more
surfels by generating a cloud of new points and taking the new (blue) points onto the
surface implied by the input (white) surfel cloud. Below, PointShop’s renderings of the
original and the filled-in set of surfels.
A dense, uniform point sample from a natural surface inherently
contains good normal information, so using surfels is not really
helpful with such data. To check the efficiency of our implementation, however, we used it to fill in the surface of a densely sampled
model, a typical application for point-set surfaces. The results are
shown in Figure 10. We found that even though we use a non-linear
optimization, we are less than twice as slow as the MLS projection
heuristic implemented in the ScanTools plugin in PointShop. We
took a cloud of 77,428 surfels onto the extremal surface implied
by an input cloud, also consisting of of 77,428 surfels, in 16 seconds, while PointShop’s procedure (also without using the second
polynomial approximation step) required 9 seconds.
8
Discussion
Many questions about the MLS surface and other extremal surfaces
remain.
In practice, the main question is which extremal surfaces, if any,
are good choices for solving specific modeling problems. For instance, surface reconstruction from noisy laser range data seems to
require a surface definition which incorporates information such as
normals, scan direction, and a good error model, with less emphasis on performance. Is there a good extremal surface solution, and
what computational methods are appropriate in this case?
There does not yet seem to be an ideal way to find the vector field
n induced by an input point cloud. The vector field n produced by
the MLS energy function has singularities quite close to the point
cloud, which makes it difficult to work with. One possibility might
be to assign approximate normals at the input points using MLS,
and then use the resulting surfels as input to our point-set surface
Acknowledgments
We thank David Levin for corresponding with us on these questions. We thank an anonymous referee for suggesting the second
half of the proof of Claim 1, and Peter Schröder and an anonymous
referee for suggestions on the presentation. We are grateful to the
team at ETH Zurich who developed and published PointShop 3D.
We thank Holly Rushmeier and the IBM TJ Watson Research Center for the use of the ram model.
References
A DAMSON , A., AND A LEXA , M. 2003. Approximating and intersecting surfaces from points. In Proceedings of EG Symposium
on Geometry Processing 2003, 245–254.
A DAMSON , A., AND A LEXA , M. 2003. Ray tracing point set
surfaces. In Proceedings of Shape Modeling International 2003,
272–279.
A LEXA , M., B EHR , J., C OHEN -O R , D., F LEISHMAN , S., L EVIN ,
D., AND S ILVA , C. T. 2003. Computing and rendering point
set surfaces. IEEE Transactions on Visualization and Computer
Graphics 9, 1, 3–15. An earlier version appeared in IEEE Visualization 2001.
C ARR , J., B EATSON , R., C HERRIE , J., M ITCHELL , T., F RIGHT,
W., M C C ALLUM , B., AND E VANS , T. 2001. Reconstruction
and representation of 3d objects with radial basis functions. ACM
SIGGRAPH ’01, 67–76.
C URLESS , B., AND L EVOY, M. 1996. A volumetric method for
building complex models from range images. ACM SIGGRAPH
‘96, 303–312.
E DELSBRUNNER , H., AND H ARER , J. to appear. Jacobi sets
of multiple morse functions. In Foundations of Computational
Mathematics, F. Cucker, Ed. Cambridge University Press.
F LEISHMAN , S., C OHEN -O R , D., A LEXA , M., AND S ILVA , C. T.
2003. Progressive point set surfaces. ACM Transactions on
Graphics 22, 4, 997–1011.
r
G UO , X., AND Q UIN , H. 2003. Dynamic sculpting and deformation of point set surfaces. In 11th Pacific Conference on Computer Graphics and Applications (PG’03), 123–130.
G UY, G., AND M EDIONI , G. 1997. Inference of surfaces, 3d
curves and junctions from sparse, noisy, 3d data. IEEE Trans. on
Pattern Analysis and Machine Intelligence 19, 11, 1265–1277.
L EVIN , D. 2003. Mesh-independent surface interpolation. In
Geometric Modeling for Scientific Visualization, G. Brunnett,
B. Hamann, K. Mueller, and L. Linsen, Eds. Springer-Verlag.
M EDEROS , B., V ELHO , L., AND DE F IGUEIREDO , L. H. 2003.
Moving least squares multiresolution surface approximation. In
Proceedings of SIBGRAPI 2003 - XVI Brazilian Symposium on
Computer Graphics and Image Processing.
M EDIONI , G., L EE , M.-S., AND TANG , C.-K. 2000. A Computational Framework for Segmentation and Grouping. Elsevier.
PAULY, M., K EISER , R., KOBBELT, L., AND G ROSS , M. 2003.
Shape modeling with point-sampled geometry. ACM SIGGRAPH 2003, 641–650.
PAULY, M. 2003. Point Primitives for Interactive Modeling and
Processing of 3D Geometry. PhD thesis, ETH Zurich.
P RESS , W., F LANNERY, B., T EUKOLSKY, S., AND V ETTERLING ,
W. 1992. Numerical Recipes in C, 2nd ed. Cambridge University
Press.
y
S
f(r)
f(y)
Figure 11:
Function f is described in Section 2. We note here that it is generally
the case that f (y) 6= f (r) for a point y on the line segment connecting r with f (r).
course the energy value e(x, a) is the same in both Jy and Jr . The
problem is that e(x, a) is not necessarily a minimum in Jy ; the neighborhood of (x, a) in Jy is different from the neighborhood of (x, a)
in Jr , and elements (x0 , a0 ) in the neighborhood of (x, a) in Jy , not
belonging to Jr , may (and generally do) have a lower values of the
energy function e. We cannot claim, “e(x, a) is a minimum in Jr ,
(x, a) belongs to Jy , therefore e(x, a) is a minimum in Jy .”
In PointShop, Pauly [Pauly 2003] describes different iterative
procedure, illustrated in Figure 12. In this procedure, we begin with
r
Hx
x1
c1
X IE , H., WANG , J., H UA , J., Q UIN , H., AND K AUFMAN , A.
2003. Piecewise c1 continuous surface reconstruction of noisy
point clouds via local implicit quadric regression. IEEE Visualization 2003, 91–98.
Z WICKER , M., PAULY, M., K NOLL , O., AND G ROSS , M. 2002.
Pointshop 3d: An interactive system for point-based surface editing. ACM SIGGRAPH 2002, 322–329.
A
Appendix - Projection procedures
Interestingly, two procedures described in the literature for producing points on the MLS surface both output points very near, but not
actually on, the surface, for almost all inputs. While these procedures obviously work well in practice, it seems important to recognize these subtleties in the effort to develop a good theory of MLS
and other extremal surfaces. In this appendix we discuss the technical problems with these procedures. As a concrete way of showing
what goes wrong, we include a Mathematica file as supplementary
material with this paper, giving an example of applying both procedures to a generic point r in space and observing that indeed the
resulting point x does not meet the definition of a point on the MLS
surface.
Since the stationary points of Levin’s projection function f , described in Section 2 are defined to be the points of the MLS surface, we know that if we iterate the procedure and find that it converges to a point x, then x ∈ S. An early argument of Levin’s (in
the manuscript version of his paper) suggested that one iteration
suffices, that is that f (r) ∈ S for any r. While this seems very plausible, in fact there is a subtle problem, specifically in the following
proposition, which turns out to be false: that for any y on the line
segment connecting r and x = f (r), we also have f (y) = x. The
actual situation is shown in Figure 11.
Recall that in the projection procedure we find x by finding a
pair (x, a) which is a local minimum of e over the set Jr . The statement that f (r) = x seems plausible because if we want to compute
f (y) we consider the set Jy , and (x, a) certainly belongs to Jy and of
1
x2
Figure 12: Pauly’s procedure: the points of P are weighted by a
guess xi , and the total-least-squares best-fitting plane Hxi , passing
through the weighted centroid c, is computed. The new guess x i+1
is the projection of r onto Hxi .
an estimated projection point x1 for r. We use x1 to assign weights
θ (x, pi ) to the points pi of the input point cloud P. We then find
the total-least-squares best-fit plane Hx1 to the weighted set P; notice that although Hx1 passes through the centroid c of the weighted
point cloud P, in general it does not pass through x1 . We project r
onto Hx1 , giving a new estimate of the projection x2 , and we iterate.
If this procedure converges (which it does very reliably), we output
the resulting point x. This procedure has the distinct advantage of
requiring no non-linear optimization.
For most r, however, the output point x is not a point of the MLS
surface S. Again, it is tempting to think so: at convergence x ∈ Hx ,
and the normal to Hx is indeed n(x) as defined by the MLS energy
function. And since Hx is the total-least-squares best-fit plane, x is
a local minimum, along the line `x,n(x) , of the energy function
e0 (y, a) = ∑(ha, pi i − ha, yi)2 θ (x, pi )
i
Notice that the weights on the points are fixed in e0 , while for
eMLS they vary with y; so e0 is not precisely the same function
as eMLS . It is certainly true that at the point of convergence x,
e0 (x, a) = eMLS (x, a). But we cannot claim that, “Since (x, a) is
a local minimum of e0 (y, n(x)) on `x,n(x) , and e0 (x, a) = eMLS (x, a),
therefore (x, a) is a local minimum of eMLS (y, n(x)) on `x,a .” The
two different functions generally have very slightly different minima.
Download